Which of the following best describes transfer learning in practice?

• A. It eliminates the need for any data in the target domain.
• B. It requires training from scratch on full data.
• C. It reuses a model trained in a related domain with fine-tuning when target data is scarce.
• D. It guarantees higher accuracy without any tuning.

Answer: (C)

Transfer learning in practice means taking a model that’s already trained on a related domain and adapting it to a new target task, usually by fine-tuning when the target data are scarce. The big win here is that the model has already learned useful representations from a large, related dataset, so you don’t start from scratch. In practice you can use the pre-trained model as a fixed feature extractor or adjust some or all of its layers with a smaller learning rate on your target data. This approach is especially powerful when labeled data in the target domain are limited, because it leverages the learned patterns to generalize better with less new data. For example, a vision model pretrained on a broad image dataset can be fine-tuned on a small set of branded product images to classify ad creatives, or a language model trained on general text can be fine-tuned for a brand’s customer queries with only a modest amount of task-specific data.

It’s not about removing data in the target domain outright; you still need some data to tailor the model to the new task. Training from scratch on full data ignores the benefits of existing learned representations. And while transfer learning can boost performance, it doesn’t guarantee higher accuracy without careful tuning—there may be a need to adjust which layers are trainable, regularization, and learning rates to suit the target domain.